The paper "Deep Learning-based Fruit Freshness Classification and Detection with CMOS Image Sensors and Edge Processors" explores advancements and applications of AI in agricultural imaging, specifically for fruit freshness classification. High-resolution and high-speed CMOS sensors, such as ON Semiconductor’s 13 MP AR1335 and XGS-12, play a critical role in capturing the quality images necessary for effective CNN-based classification and detection tasks. These sensors, integrated with edge computing platforms, provide the requisite image quality and processing capability needed for real-time AI applications.
The study emphasizes the use of MobileNetV2 for its balance between accuracy, latency, and the number of parameters, making it suitable for edge computing. MobileNetV2, featuring depthwise-separable convolutions and inverted residual structures, was chosen for its efficiency. The model was initialized with pretrained ImageNet weights and fine-tuned on a custom dataset of 30,000 images to classify fresh and rotten apples, bananas, and oranges, achieving a remarkable 97% accuracy. This approach leverages the strengths of transfer learning, enhancing the model's performance with a relatively small number of training parameters.
For object detection, the paper employs SSD (Single Shot Detector) with MobileNetV2 as the backbone, facilitating real-time detection and classification tasks. The choice of SSD over region proposal network-based architectures allows for faster and more efficient processing, crucial for edge platforms like NVIDIA Jetson Xavier. The detection model, however, shows a lower accuracy compared to the classification model, highlighting the need for further improvements and more extensive training data.
Training data included images from both a custom-captured dataset and a publicly available Kaggle fruits dataset, augmented with various transformations such as scaling, translation, rotation, and brightness variation to improve model robustness. The combination of high-quality images and diverse augmentations ensures the CNN models are well-trained and capable of handling real-world variations in the data.
Performance evaluation underscores the importance of high-quality images and optimized CNN architectures for achieving accurate real-time performance on edge platforms. While the image classification model achieved high accuracy, the object detection model showed potential for improvement. Ongoing efforts to expand the dataset and improve annotations are aimed at enhancing the accuracy and robustness of the object detection model, ensuring better real-time performance in practical applications.
(2020 Tejaswini Ananthanarayana1, Raymond Ptucha, Sean C. Kelly)
